Papers by Zenan Zhou
FB-Bench: A Fine-Grained Multi-Task Benchmark for Evaluating LLMs’ Responsiveness to Human Feedback (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing research focuses on benchmarking LLMs in single-turn dialogues, neglecting the nuanced nature of human feedback within real-world usage scenarios. |
| Approach: | They propose a fine-grained, multi-task benchmark designed to evaluate LLMs’ responsiveness to human feedback under real-world usage scenarios in Chinese. |
| Outcome: | The proposed benchmarks show that human feedback can significantly impact LLMs’ responsiveness in real-world usage scenarios. |
Reinforcement Learning on Pre-Training Data (2026.acl-long)
Copied to clipboard
Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, Kun Li, Haoyuan Wu, null Wujiajia, Zihao Zheng, Chenchen Zhang, Kun Shi, Xue Gong, Qi Yi, Ruibin Xiong, Tingqiang Xu, Yuhao Jiang, Jianfeng Yan, Yuyuan Zeng, Guanghui Xu, Jinbao Xue, Zhijiang xu, Zheng Fang, Shuai LI, Qibin Liu, Xiaoxue Li, Zhuoyu Li, Yangyu Tao, Fei Gao, Cheng Jiang, Bochao Wang, Kai Liu, Jianchen Zhu, Wai Lam, Bo Zhou, Di Wang
| Challenge: | Recent progress in large language models is driven by scaling of training compute through pre-training with nexttoken prediction (NTP) or post-training (RL) Pre-training using NTP enables models to acquire extensive knowledge and skills from general data, but it suffers from data inefficiency and catastrophic forgetting in continual learning settings. |
| Approach: | They propose to scale training compute through pre-training with next-token prediction (NTP) or post-training by scaling reinforcement learning (RL) to improve learning from general data. |
| Outcome: | Experiments on multiple benchmarks and models show that the proposed approach improves continual pre-training and provides a strong foundation for post-training on Qwen3-8B-Base. |
Reasoning Over Semantic-Level Graph for Fact Checking (2020.acl-main)
Copied to clipboard
| Challenge: | Existing methods for fact checking use string concatenation or fusing features of isolated evidence sentences. |
| Approach: | They propose a method suitable for reasoning about the semantic-level structure of evidence . they use graph convolutional network and graph attention network to exploit the structure . |
| Outcome: | The proposed method improves claim verification accuracy and FEVER score on a benchmark dataset. |
Discovery and Reinforcement of Tool-Integrated Reasoning Chains via Rollout Trees (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to augment Large Language Models (LLMs) with computational capabilities have focused on short Chain-of-thought (CoT) integrating tool-use into long CoT remains underexplored due to the scarcity of training data and the challenge of integrating it without compromising the model’s intrinsic long-chain reasoning. |
| Approach: | They propose a framework that enables spontaneous tool-use during long CoT reasoning without additional human annotation. |
| Outcome: | Experiments on AIME and GPQA-Diamond show that DART significantly outperforms existing methods, successfully harmonizing tool execution with long CoT reasoning. |
CFBench: A Comprehensive Constraints-Following Benchmark for LLMs (2025.acl-long)
Copied to clipboard
Tao Zhang, ChengLIn Zhu, Yanjun Shen, Wenjing Luo, Yan Zhang, Hao Liang, Tao Zhang, Fan Yang, Mingan Lin, Yujing Qiao, Weipeng Chen, Bin Cui, Wentao Zhang, Zenan Zhou
| Challenge: | Existing evaluations of Large Language Models (LLMs) focus on fragmented constraints or narrow scenarios, but they overlook the comprehensiveness and authenticity of constraints from the user’s perspective. |
| Approach: | They propose a Chinese Comprehensive Constraints Following Benchmark for LLMs that compiles constraints from real-world instructions and constructs a systematic framework for constraint types. |
| Outcome: | The proposed framework integrates multi-dimensional assessment criteria with requirement prioritization, covering various perspectives of constraints, instructions, and requirement fulfillment. |
Analytical Reasoning of Text (2022.findings-naacl)
Copied to clipboard
Wanjun Zhong, Siyuan Wang, Duyu Tang, Zenan Xu, Daya Guo, Yining Chen, Jiahai Wang, Jian Yin, Ming Zhou, Nan Duan
| Challenge: | Existing models with implicit reasoning ability struggle to solve analytical reasoning of text. |
| Approach: | They propose an approach to analyze text and use it to perform reasoning over it. |
| Outcome: | The proposed approach outperforms pre-trained models on an analysis of the Law School Admission Test dataset. |
Neural Deepfake Detection with Factual Structure of Text (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to deepfake detection typically represent documents with coarse-grained representations, but they struggle to capture factual structures of documents. |
| Approach: | They propose a graph-based model that captures factual structures of documents for deepfake detection. |
| Outcome: | The proposed model improves strong base models built with RoBERTa on two public deepfake datasets. |
MM-Verify: Enhancing Multimodal Reasoning with Chain-of-Thought Verification (2025.acl-long)
Copied to clipboard
| Challenge: | MM-Verifier and MM Reasoner are a powerful multimodal reasoning model . large language models (LLMs) have demonstrated exceptional performance across tasks spanning myriad domains. |
| Approach: | They propose a method which combines tree search and verification to generate high-quality chain-of-thought data. |
| Outcome: | The proposed method outperforms all larger models on the MathCheck, MathVista, and MathVerse benchmarks. |
Rhombus: Incentivizing Coordination in Parallel Thinking through Reinforcement Learning (2026.findings-acl)
Copied to clipboard
Ziyuan Nan, Qi Yi, Di Huang, Yutong Wu, Guanhua Huang, Xue Gong, Kejiao Li, Yuhao Jiang, Chenchen Zhang, Zenan Xu, Xing Hu, Bo Zhou
| Challenge: | Parallel thinking is a promising avenue for scaling test-time compute in Large Language Models . however, coordinating the exploration and aggregation stages remains challenging . |
| Approach: | They propose a parallel thinking framework that explicitly incentivizes coordination between components via end-to-end reinforcement learning. |
| Outcome: | The proposed framework improves accuracy by 6.0% over long chain-of-thought baselines while reducing wall-clock latency by 39.4% under matched token budgets. |